search

openai / guided-diffusion

MIT License
6.06k stars 807 forks source link

FID score on Imagenet-256 dataset #14

Closed cientgu closed 2 years ago

cientgu commented 2 years ago

It's really a great work! However, on ImageNet-256 dataset, I test the FID between 50k images from the validation set and 50k images from the training set, the FID is about 7.6. (using two popular FID code, https://github.com/mseitzer/pytorch-fid and https://github.com/tsc2017/Frechet-Inception-Distance , which get very similar results). So I am really confused how to achieve the FID score 4.59 in your paper. Could you share the code to calculate the FID score? I just want to make sure my way to calculate FID is not wrong.

je-santos commented 1 year ago

I also got ~7.5 using 10k images. @cientgu any luck?

Yuxin-Zhang-Jasmine commented 1 year ago

There is an evaluator.py in the folder evaluations